Application Research of Model-Free Reinforcement Learning under the Condition of Conditional Transfer Function with Coupling Factors

Abstract

Dynamic systems are ubiquitous in nature. The analysis of the stability and performance of dynamic systems has been a research hotspot in control science and operations research for a long time. In this paper, we construct and analyze an actual sequential decision-making problem of dynamic system. The Model-Free reinforcement learning algorithms are used to optimize this problem. The problem is analyzed in detail through adaptive control theory and information theory, also the extreme performance of the algorithm is pointed out. In this paper, we select three classic Model-Free reinforcement learning algorithms, DQN, DQN-PER, and PPO, to compare and analyze their performance on the timing series decision problem we construct.